data

How to generate a dataset from videos

0. Record videos of the visual environment, yielding mp4 files
1. Generate images: ./gen_imgs_from_vids.sh images/out/path/ wildcard/path/to/videos/*.mp4
2. Turn images into black and white, scale them to 160x120, then apply edge detection if anything but CORF is used: prepare_imgs.py corf|sobel wildcard/to/images*.jpg path/to/output/. If you choose NOT to perform edge detection on your images at all, just select corf, and skip step 3.
3. If you chose CORF as your edge detection algorithm in the previous step, run Matlab script convertAllInFolder2DPar.m in parallel under matlab/faster_corf/ to extract contour (see instruction in script), else just continue on with step 4
4. Enrich dataset with mirror images if applicable: run mirror_imgs.py wildcard/to/images*.png
5. Merge images into hdf5 file: merge_imgs.py path/to/contour/imgs*.jpg output/path/something.hdf5

The resulting hdf5 file should contain separate randomized training and test sets, by a default 90-10% fold (specify in merge_imgs.py if needed).

Datasets used in the study are also available. Download, then place them in the data folder:

• hand gesture dataset,
• table / reaching movement dataset.